The button is ready.
One tap and the post is shared.
Then the interface interrupts.
“Before you share, would you like to read the article?”
For perhaps two seconds, an effortless action becomes slightly less effortless.
That tiny delay is friction.
Whether it improves judgment is one of the most practically important questions in contemporary information design.
What counts as friction?
In this context, friction means a designed increase in the effort, time, reflection or number of steps required before an information-related action is completed.
Examples include:
- an accuracy prompt before sharing;
- a reminder that a headline has not been opened;
- a requirement to explain why a claim seems true or false;
- a short delay before forwarding content;
- a warning or contextual label;
- a prompt to inspect a source;
- a confidence rating before an answer is submitted;
- a cognitive forcing function that delays acceptance of an AI recommendation.
These interventions differ in mechanism.
Some redirect attention.
Some add time.
Some increase accountability.
Some reveal missing information.
Some merely make an action harder.
Treating all of them as one thing obscures the evidence.
The case for friction: attention is often part of the problem
A major line of misinformation research argues that people sometimes share low-quality content not because they fundamentally prefer falsehood but because accuracy is not salient at the moment of sharing.
Pennycook and colleagues reported that prompting people to consider accuracy improved the quality of subsequent sharing intentions in studied social-media contexts.
A 2022 Nature Communications paper found accuracy prompts to be replicable across multiple experiments and contexts, while also documenting heterogeneity.
The theoretical appeal is straightforward.
If an interface rewards speed, novelty and reaction, a small intervention can change what the person is attending to just before action.
Friction, in this view, is not teaching a new capacity in real time.
It is giving an existing capacity a chance to enter the decision.
Explanation prompts offer another route
Lisa Fazio tested a more effortful intervention.
Participants were asked to pause and consider why a headline might be true or false.
In her 2020 experiment with 501 MTurk participants, this explanatory pause reduced stated intentions to share false headlines.
That finding supports the broader idea that slowing down can matter when the added step directs attention toward evidence.
But it does not imply that any delay works.
A two-second spinner has no obvious epistemic content.
The design of the friction matters.
Friction can act as a cognitive forcing function
Human-AI interaction research provides a parallel literature.
When an AI system presents a recommendation fluently and immediately, users may accept it with insufficient independent evaluation.
Cognitive forcing functions can require the human to make an initial judgment, inspect evidence or otherwise engage before seeing or accepting the AI answer.
Buçinca and colleagues found that such interventions could reduce overreliance on AI in experimental tasks, but with costs to user experience and effort.
This trade-off is central.
Friction may protect deliberation precisely by making a task less convenient.
The question is whether the protection is worth the cost.
The case against simple friction stories: sharing can be habitual
Not all misinformation sharing is a momentary failure to think.
Ceylan, Anderson and Wood found evidence that social-media sharing can be strongly habitual. In their PNAS work, highly habitual users were less responsive to informational signals that would ordinarily shape deliberate sharing.
This matters because a prompt designed for deliberation may have limited effect when behavior is strongly cued by habit and reward structure.
The same person can know that accuracy matters and still enact a practiced sharing routine.
Friction that does not alter the underlying cue-reward architecture may therefore produce only modest change.
Effects can differ across people and contexts
Accuracy-prompt research has generated debate about heterogeneity, including partisan differences and task sensitivity.
A 2024 adversarial collaboration led by Martel and colleagues examined whether accuracy prompts work similarly across partisan groups. The value of adversarial collaboration here is methodological: researchers with different expectations agree on tests and analyses before seeing the result.
The broader lesson is that mean effects should not be treated as universal effects.
An intervention can improve average discernment while helping some people more than others, doing little for habitual sharers, or changing response bias alongside sensitivity.
Design decisions should therefore consider distributional effects, not only average effect size.
Too much friction may become noise
A warning shown once can attract attention.
A warning shown constantly can become part of the wallpaper.
This is habituation.
People learn interfaces.
They learn where the dismiss button is. They learn which prompts can be clicked through. They may also develop generalized distrust if warnings are broad or poorly targeted.
Friction can therefore lose effectiveness with repetition.
Long-term field evidence is especially important because laboratory participants encounter an intervention briefly, while real users may encounter it hundreds of times.
Friction has accessibility and equity costs
Extra steps are not neutral.
They consume time, working memory, motor effort and language comprehension.
For users with cognitive disabilities, low literacy, fatigue, attention difficulties or limited connectivity, an additional verification step may impose disproportionate cost.
Professionals in high-volume environments may also face cumulative burden.
A hospital, newsroom or emergency system cannot simply add reflection prompts everywhere without considering workflow consequences.
Good friction should therefore be selective, comprehensible and proportionate to risk.
Accessibility is part of epistemic design, not an afterthought.
The Copenhagen friction model adds a useful warning
A 2025 perspective in npj Complexity by Jahn and colleagues, including researchers in Copenhagen, uses modeling to explore friction interventions against misinformation.
Its status needs to be described carefully.
The work is not a large completed field trial proving that a specific platform intervention improves real-world information quality.
It uses modeling and proposes empirical directions.
The model suggests an important possibility: friction can reduce overall sharing volume without necessarily improving the quality of what remains, and excessive friction may become counterproductive unless combined with learning or adaptation.
This is exactly the kind of result that should inform design without being overstated as settled causal evidence.
What outcome are we trying to improve?
“Better judgment” is too broad.
A friction intervention might affect:
- belief accuracy;
- sharing discernment;
- response time;
- source inspection;
- confidence calibration;
- overreliance on AI;
- total sharing volume;
- trust in the platform;
- user satisfaction;
- learning retention;
- accessibility;
- false-positive skepticism.
An intervention can improve one outcome while worsening another.
For example, a warning might reduce false sharing but also reduce true sharing.
A forced delay might improve independent reasoning but make the system unusable in time-sensitive settings.
A confidence prompt might reveal uncertainty while also changing the underlying decision process.
Evaluation should therefore specify the intended benefit and the possible collateral cost.
Friction can target attention, habit or accountability
A useful way to organize the field is by mechanism.
Attention friction makes accuracy or source quality salient.
Temporal friction inserts delay.
Epistemic friction requires a reason, source check or confidence judgment.
Behavioral friction adds steps that interrupt habitual action.
Accountability friction makes the user anticipate that a decision may need explanation.
AI forcing functions preserve a stage of independent human judgment before automated advice is accepted.
These mechanisms should not be expected to produce identical effects.
A design that fails may be targeting the wrong mechanism.
The strongest position is conditional
The evidence does not support the claim that friction is always good.
It also does not support the claim that friction is merely paternalistic obstruction.
The evidence supports a more conditional conclusion.
Friction can improve some forms of judgment when it is well targeted, when the relevant capacity already exists, when the cost is proportionate and when the intervention directs attention toward diagnostic information.
Its benefits are less certain when behavior is strongly habitual, when prompts are repeated until ignored, when tasks are low stakes, when accessibility costs are high or when the added step does not actually improve the information available.
Timing may matter as much as intensity
Friction placed too early can obstruct harmless exploration.
Placed too late, it may arrive after commitment has already formed.
A prompt shown before a person reads a headline can affect attention differently from a prompt shown at the moment of sharing. A verification step before an AI recommendation is revealed can preserve independent judgment differently from a warning shown after the recommendation has anchored the user.
This suggests a design principle: locate friction near the decision point where the relevant error is generated.
For impulsive sharing, the share action may be the target. For automation bias, the target may be before acceptance of the recommendation. For source evaluation, the target may be when provenance becomes uncertain.
The same intervention can therefore succeed or fail depending on placement.
Reversibility should influence how much friction is justified
Some decisions are cheap to undo.
Others are not.
A typo in a private note can be corrected. Sending sensitive data to the wrong system may not be reversible. Sharing a false accusation about a person can cause damage long before a correction reaches the same audience.
High irreversibility strengthens the case for friction.
This helps explain why financial transfers, clinical procedures and security systems already contain confirmation steps while low-risk browsing usually does not.
Epistemic systems can use the same logic without copying it mechanically.
The expected harm of error, not a generalized preference for slowness, should guide intervention intensity.
Friction can become performative compliance
Users learn to satisfy prompts without engaging with them.
They tick boxes, type minimal explanations and click “I understand.”
This creates the appearance of deliberation without its substance.
Evaluation should therefore measure whether the mechanism changed, not merely whether the prompt was displayed.
Did users inspect more sources? Did they distinguish true from false items better? Did confidence track accuracy more closely? Did overreliance on AI fall? Did the effect persist after repeated exposure?
A compliance metric such as “percentage of users who clicked through the warning” may say almost nothing about judgment quality.
Adaptive friction is promising and ethically complicated
One response to prompt fatigue is to add friction selectively.
Systems could trigger extra verification for high-risk actions, unfamiliar sources, anomalous content or low-confidence model outputs.
This may reduce burden, but it introduces new questions.
Who decides what counts as risky? What data are used to infer that risk? Could selective prompts disadvantage particular groups or viewpoints? Can users understand why the system interrupted them?
Adaptive friction therefore requires transparency and auditing.
A system that protects judgment by secretly manipulating which users are slowed down can create governance problems even if average accuracy improves.
The comparison condition changes the interpretation
A friction intervention may look effective compared with an interface optimized for instantaneous sharing.
That does not tell us whether it is better than education, improved ranking, provenance display, reduced virality or a different default.
Policy and product decisions require comparative effectiveness.
The relevant question is often not “Does this prompt work?” but “Is this the best available way to reduce this error at this cost?”
That comparison should include non-cognitive interventions. Sometimes redesigning the information environment is more effective than asking each user to overcome it repeatedly.
Measure adaptation over time, not only first exposure
A friction intervention can look effective on day one because it is novel.
That novelty is not the same as durable mechanism.
Users may initially read a prompt carefully and later dismiss it automatically. They may discover shortcuts. Organizations may route work around a burdensome control. Conversely, repeated exposure may teach a useful habit so that the prompt becomes less necessary over time.
Longitudinal evaluation should therefore ask whether effects persist, decay, reverse or become internalized.
This is particularly important for AI interfaces, where users can interact with the same assistance pattern many times each day.
A one-session laboratory effect is evidence of possibility. It is not yet evidence of durable institutional benefit.
Friction can also protect autonomy
Friction is often described as paternalistic because it interferes with immediate choice.
But instant action is not always more autonomous.
Interfaces can exploit defaults, urgency, salience and habit to accelerate behavior before reflective preferences enter. A pause can sometimes create room for the user's longer-term intention to compete with the immediate cue.
The ethical question is therefore not simply freedom versus friction.
It is who controls the timing and for what purpose.
Friction that is transparent, proportionate and aligned with the user's stated goals can support agency. Hidden or manipulative friction can obstruct it.
The same technical mechanism can serve very different governance values.
What the evidence does not show
Current evidence does not show that:
- any delay improves judgment;
- accuracy prompts eliminate misinformation sharing;
- laboratory or survey-intention effects automatically persist on live platforms;
- more friction is better than less friction;
- the same prompt works equally across political, cultural and age groups;
- warnings improve sensitivity rather than merely shifting people toward skepticism;
- friction can replace education, trustworthy institutions or platform accountability;
- a model-based result is equivalent to a completed field experiment;
- reducing total sharing necessarily improves the truthfulness of what is shared;
- users should bear all responsibility for resisting systems designed to maximize rapid engagement.
These are design and evidence boundaries, not footnotes.
From interface to institution
A society already contains friction.
Peer review is friction. Editorial review is friction. A second signature on a financial transfer is friction. Clinical consent procedures are friction. Courts require reasons and evidence because consequential action should not always be instantaneous.
The deeper question is where friction belongs.
If institutions remove every pause in the name of convenience, they may create environments in which error travels faster than correction.
If institutions add friction everywhere, they can create bureaucracy that people learn to evade.
The design problem is therefore constitutional in miniature: where should speed be protected, and where should judgment be protected from speed?
Institute synthesis: friction as a scarce resource
The Institute proposes a bounded framing: friction is a scarce design resource.
It should be spent where the expected value of reflection exceeds the cost of interruption.
That is a synthesis, not a validated formula.
It suggests three questions for designers and institutions:
What error are we trying to prevent?
What mechanism will this friction actually engage?
What burden will it impose, especially on people already carrying more cognitive or accessibility load?
The share button is still there.
The pause is only useful if something worthwhile can happen inside it.
Research references
- Pennycook G, Epstein Z, Mosleh M, Arechar AA, Eckles D, Rand DG. Shifting attention to accuracy can reduce misinformation online. Nature. 2021;592:590-595. https://doi.org/10.1038/s41586-021-03344-2
- Pennycook G, Rand DG. Accuracy prompts are a replicable and generalizable approach for reducing the spread of misinformation. Nature Communications. 2022;13:2333. https://doi.org/10.1038/s41467-022-30073-5
- Fazio LK. Pausing to consider why a headline is true or false can help reduce the sharing of false news. Harvard Kennedy School Misinformation Review. 2020. https://doi.org/10.37016/mr-2020-009
- Ceylan G, Anderson IA, Wood W. Sharing of misinformation is habitual, not just lazy or biased. Proceedings of the National Academy of Sciences. 2023;120(4):e2216614120. https://doi.org/10.1073/pnas.2216614120
- Martel C, Rathje S, Clark CJ, Pennycook G, Van Bavel JJ, Rand DG, van der Linden S. On the Efficacy of Accuracy Prompts Across Partisan Lines: An Adversarial Collaboration. Psychological Science. 2024. https://doi.org/10.1177/09567976241232905
- Jahn L, et al. A perspective on friction interventions to curb the spread of misinformation. npj Complexity. 2025. https://doi.org/10.1038/s44260-025-00051-1
- Buçinca Z, Malaya MB, Gajos KZ. To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making. Proceedings of the ACM on Human-Computer Interaction. 2021. https://doi.org/10.1145/3449287